urcu-bp.c

   1 /*
   2  * urcu-bp.c
   3  *
   4  * Userspace RCU library, "bulletproof" version.
   5  *
   6  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   7  * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  *
  23  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  24  */
  25
  26 #define _GNU_SOURCE
  27 #define _LGPL_SOURCE
  28 #include <stdio.h>
  29 #include <pthread.h>
  30 #include <signal.h>
  31 #include <assert.h>
  32 #include <stdlib.h>
  33 #include <string.h>
  34 #include <errno.h>
  35 #include <poll.h>
  36 #include <unistd.h>
  37 #include <sys/mman.h>
  38
  39 #include "urcu/wfcqueue.h"
  40 #include "urcu/map/urcu-bp.h"
  41 #include "urcu/static/urcu-bp.h"
  42 #include "urcu-pointer.h"
  43 #include "urcu/tls-compat.h"
  44
  45 #include "urcu-die.h"
  46
  47 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  48 #undef _LGPL_SOURCE
  49 #include "urcu-bp.h"
  50 #define _LGPL_SOURCE
  51
  52 #ifndef MAP_ANONYMOUS
  53 #define MAP_ANONYMOUS MAP_ANON
  54 #endif
  55
  56 #ifdef __linux__
  57 static
  58 void *mremap_wrapper(void *old_address, size_t old_size,
  59                 size_t new_size, int flags)
  60 {
  61         return mremap(old_address, old_size, new_size, flags);
  62 }
  63 #else
  64
  65 #define MREMAP_MAYMOVE  1
  66 #define MREMAP_FIXED    2
  67
  68 /*
  69  * mremap wrapper for non-Linux systems not allowing MAYMOVE.
  70  * This is not generic.
  71 */
  72 static
  73 void *mremap_wrapper(void *old_address, size_t old_size,
  74                 size_t new_size, int flags)
  75 {
  76         assert(!(flags & MREMAP_MAYMOVE));
  77
  78         return MAP_FAILED;
  79 }
  80 #endif
  81
  82 /* Sleep delay in us */
  83 #define RCU_SLEEP_DELAY         1000
  84 #define INIT_NR_THREADS         8
  85 #define ARENA_INIT_ALLOC                \
  86         sizeof(struct registry_chunk)   \
  87         + INIT_NR_THREADS * sizeof(struct rcu_reader)
  88
  89 /*
  90  * Active attempts to check for reader Q.S. before calling sleep().
  91  */
  92 #define RCU_QS_ACTIVE_ATTEMPTS 100
  93
  94 static
  95 int rcu_bp_refcount;
  96
  97 static
  98 void __attribute__((constructor)) rcu_bp_init(void);
  99 static
 100 void __attribute__((destructor)) rcu_bp_exit(void);
 101
 102 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 103
 104 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 105 static int initialized;
 106
 107 static pthread_key_t urcu_bp_key;
 108
 109 #ifdef DEBUG_YIELD
 110 unsigned int rcu_yield_active;
 111 DEFINE_URCU_TLS(unsigned int, rcu_rand_yield);
 112 #endif
 113
 114 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
 115
 116 /*
 117  * Pointer to registry elements. Written to only by each individual reader. Read
 118  * by both the reader and the writers.
 119  */
 120 DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader);
 121
 122 static CDS_LIST_HEAD(registry);
 123
 124 struct registry_chunk {
 125         size_t data_len;                /* data length */
 126         size_t used;                    /* amount of data used */
 127         struct cds_list_head node;      /* chunk_list node */
 128         char data[];
 129 };
 130
 131 struct registry_arena {
 132         struct cds_list_head chunk_list;
 133 };
 134
 135 static struct registry_arena registry_arena = {
 136         .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
 137 };
 138
 139 /* Saved fork signal mask, protected by rcu_gp_lock */
 140 static sigset_t saved_fork_signal_mask;
 141
 142 static void mutex_lock(pthread_mutex_t *mutex)
 143 {
 144         int ret;
 145
 146 #ifndef DISTRUST_SIGNALS_EXTREME
 147         ret = pthread_mutex_lock(mutex);
 148         if (ret)
 149                 urcu_die(ret);
 150 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 151         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 152                 if (ret != EBUSY && ret != EINTR)
 153                         urcu_die(ret);
 154                 poll(NULL,0,10);
 155         }
 156 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 157 }
 158
 159 static void mutex_unlock(pthread_mutex_t *mutex)
 160 {
 161         int ret;
 162
 163         ret = pthread_mutex_unlock(mutex);
 164         if (ret)
 165                 urcu_die(ret);
 166 }
 167
 168 static void wait_for_readers(struct cds_list_head *input_readers,
 169                         struct cds_list_head *cur_snap_readers,
 170                         struct cds_list_head *qsreaders)
 171 {
 172         int wait_loops = 0;
 173         struct rcu_reader *index, *tmp;
 174
 175         /*
 176          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 177          * indicate quiescence (not nested), or observe the current
 178          * rcu_gp.ctr value.
 179          */
 180         for (;;) {
 181                 wait_loops++;
 182                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 183                         switch (rcu_reader_state(&index->ctr)) {
 184                         case RCU_READER_ACTIVE_CURRENT:
 185                                 if (cur_snap_readers) {
 186                                         cds_list_move(&index->node,
 187                                                 cur_snap_readers);
 188                                         break;
 189                                 }
 190                                 /* Fall-through */
 191                         case RCU_READER_INACTIVE:
 192                                 cds_list_move(&index->node, qsreaders);
 193                                 break;
 194                         case RCU_READER_ACTIVE_OLD:
 195                                 /*
 196                                  * Old snapshot. Leaving node in
 197                                  * input_readers will make us busy-loop
 198                                  * until the snapshot becomes current or
 199                                  * the reader becomes inactive.
 200                                  */
 201                                 break;
 202                         }
 203                 }
 204
 205                 if (cds_list_empty(input_readers)) {
 206                         break;
 207                 } else {
 208                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 209                                 usleep(RCU_SLEEP_DELAY);
 210                         else
 211                                 caa_cpu_relax();
 212                 }
 213         }
 214 }
 215
 216 void synchronize_rcu(void)
 217 {
 218         CDS_LIST_HEAD(cur_snap_readers);
 219         CDS_LIST_HEAD(qsreaders);
 220         sigset_t newmask, oldmask;
 221         int ret;
 222
 223         ret = sigfillset(&newmask);
 224         assert(!ret);
 225         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 226         assert(!ret);
 227
 228         mutex_lock(&rcu_gp_lock);
 229
 230         if (cds_list_empty(&registry))
 231                 goto out;
 232
 233         /* All threads should read qparity before accessing data structure
 234          * where new ptr points to. */
 235         /* Write new ptr before changing the qparity */
 236         cmm_smp_mb();
 237
 238         /*
 239          * Wait for readers to observe original parity or be quiescent.
 240          */
 241         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 242
 243         /*
 244          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 245          * model easier to understand. It does not have a big performance impact
 246          * anyway, given this is the write-side.
 247          */
 248         cmm_smp_mb();
 249
 250         /* Switch parity: 0 -> 1, 1 -> 0 */
 251         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
 252
 253         /*
 254          * Must commit qparity update to memory before waiting for other parity
 255          * quiescent state. Failure to do so could result in the writer waiting
 256          * forever while new readers are always accessing data (no progress).
 257          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 258          */
 259
 260         /*
 261          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 262          * model easier to understand. It does not have a big performance impact
 263          * anyway, given this is the write-side.
 264          */
 265         cmm_smp_mb();
 266
 267         /*
 268          * Wait for readers to observe new parity or be quiescent.
 269          */
 270         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 271
 272         /*
 273          * Put quiescent reader list back into registry.
 274          */
 275         cds_list_splice(&qsreaders, &registry);
 276
 277         /*
 278          * Finish waiting for reader threads before letting the old ptr being
 279          * freed.
 280          */
 281         cmm_smp_mb();
 282 out:
 283         mutex_unlock(&rcu_gp_lock);
 284         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 285         assert(!ret);
 286 }
 287
 288 /*
 289  * library wrappers to be used by non-LGPL compatible source code.
 290  */
 291
 292 void rcu_read_lock(void)
 293 {
 294         _rcu_read_lock();
 295 }
 296
 297 void rcu_read_unlock(void)
 298 {
 299         _rcu_read_unlock();
 300 }
 301
 302 int rcu_read_ongoing(void)
 303 {
 304         return _rcu_read_ongoing();
 305 }
 306
 307 /*
 308  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 309  * Else, try expanding the last chunk. If this fails, allocate a new
 310  * chunk twice as big as the last chunk.
 311  * Memory used by chunks _never_ moves. A chunk could theoretically be
 312  * freed when all "used" slots are released, but we don't do it at this
 313  * point.
 314  */
 315 static
 316 void expand_arena(struct registry_arena *arena)
 317 {
 318         struct registry_chunk *new_chunk, *last_chunk;
 319         size_t old_chunk_len, new_chunk_len;
 320
 321         /* No chunk. */
 322         if (cds_list_empty(&arena->chunk_list)) {
 323                 assert(ARENA_INIT_ALLOC >=
 324                         sizeof(struct registry_chunk)
 325                         + sizeof(struct rcu_reader));
 326                 new_chunk_len = ARENA_INIT_ALLOC;
 327                 new_chunk = mmap(NULL, new_chunk_len,
 328                         PROT_READ | PROT_WRITE,
 329                         MAP_ANONYMOUS | MAP_PRIVATE,
 330                         -1, 0);
 331                 if (new_chunk == MAP_FAILED)
 332                         abort();
 333                 bzero(new_chunk, new_chunk_len);
 334                 new_chunk->data_len =
 335                         new_chunk_len - sizeof(struct registry_chunk);
 336                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 337                 return;         /* We're done. */
 338         }
 339
 340         /* Try expanding last chunk. */
 341         last_chunk = cds_list_entry(arena->chunk_list.prev,
 342                 struct registry_chunk, node);
 343         old_chunk_len =
 344                 last_chunk->data_len + sizeof(struct registry_chunk);
 345         new_chunk_len = old_chunk_len << 1;
 346
 347         /* Don't allow memory mapping to move, just expand. */
 348         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 349                 new_chunk_len, 0);
 350         if (new_chunk != MAP_FAILED) {
 351                 /* Should not have moved. */
 352                 assert(new_chunk == last_chunk);
 353                 bzero((char *) last_chunk + old_chunk_len,
 354                         new_chunk_len - old_chunk_len);
 355                 last_chunk->data_len =
 356                         new_chunk_len - sizeof(struct registry_chunk);
 357                 return;         /* We're done. */
 358         }
 359
 360         /* Remap did not succeed, we need to add a new chunk. */
 361         new_chunk = mmap(NULL, new_chunk_len,
 362                 PROT_READ | PROT_WRITE,
 363                 MAP_ANONYMOUS | MAP_PRIVATE,
 364                 -1, 0);
 365         if (new_chunk == MAP_FAILED)
 366                 abort();
 367         bzero(new_chunk, new_chunk_len);
 368         new_chunk->data_len =
 369                 new_chunk_len - sizeof(struct registry_chunk);
 370         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 371 }
 372
 373 static
 374 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 375 {
 376         struct registry_chunk *chunk;
 377         struct rcu_reader *rcu_reader_reg;
 378         int expand_done = 0;    /* Only allow to expand once per alloc */
 379         size_t len = sizeof(struct rcu_reader);
 380
 381 retry:
 382         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 383                 if (chunk->data_len - chunk->used < len)
 384                         continue;
 385                 /* Find spot */
 386                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 387                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 388                                 rcu_reader_reg++) {
 389                         if (!rcu_reader_reg->alloc) {
 390                                 rcu_reader_reg->alloc = 1;
 391                                 chunk->used += len;
 392                                 return rcu_reader_reg;
 393                         }
 394                 }
 395         }
 396
 397         if (!expand_done) {
 398                 expand_arena(arena);
 399                 expand_done = 1;
 400                 goto retry;
 401         }
 402
 403         return NULL;
 404 }
 405
 406 /* Called with signals off and mutex locked */
 407 static
 408 void add_thread(void)
 409 {
 410         struct rcu_reader *rcu_reader_reg;
 411         int ret;
 412
 413         rcu_reader_reg = arena_alloc(&registry_arena);
 414         if (!rcu_reader_reg)
 415                 abort();
 416         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 417         if (ret)
 418                 abort();
 419
 420         /* Add to registry */
 421         rcu_reader_reg->tid = pthread_self();
 422         assert(rcu_reader_reg->ctr == 0);
 423         cds_list_add(&rcu_reader_reg->node, &registry);
 424         /*
 425          * Reader threads are pointing to the reader registry. This is
 426          * why its memory should never be relocated.
 427          */
 428         URCU_TLS(rcu_reader) = rcu_reader_reg;
 429 }
 430
 431 /* Called with mutex locked */
 432 static
 433 void cleanup_thread(struct registry_chunk *chunk,
 434                 struct rcu_reader *rcu_reader_reg)
 435 {
 436         rcu_reader_reg->ctr = 0;
 437         cds_list_del(&rcu_reader_reg->node);
 438         rcu_reader_reg->tid = 0;
 439         rcu_reader_reg->alloc = 0;
 440         chunk->used -= sizeof(struct rcu_reader);
 441 }
 442
 443 static
 444 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 445 {
 446         struct registry_chunk *chunk;
 447
 448         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 449                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 450                         continue;
 451                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 452                         continue;
 453                 return chunk;
 454         }
 455         return NULL;
 456 }
 457
 458 /* Called with signals off and mutex locked */
 459 static
 460 void remove_thread(struct rcu_reader *rcu_reader_reg)
 461 {
 462         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 463         URCU_TLS(rcu_reader) = NULL;
 464 }
 465
 466 /* Disable signals, take mutex, add to registry */
 467 void rcu_bp_register(void)
 468 {
 469         sigset_t newmask, oldmask;
 470         int ret;
 471
 472         ret = sigfillset(&newmask);
 473         if (ret)
 474                 abort();
 475         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 476         if (ret)
 477                 abort();
 478
 479         /*
 480          * Check if a signal concurrently registered our thread since
 481          * the check in rcu_read_lock().
 482          */
 483         if (URCU_TLS(rcu_reader))
 484                 goto end;
 485
 486         /*
 487          * Take care of early registration before urcu_bp constructor.
 488          */
 489         rcu_bp_init();
 490
 491         mutex_lock(&rcu_gp_lock);
 492         add_thread();
 493         mutex_unlock(&rcu_gp_lock);
 494 end:
 495         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 496         if (ret)
 497                 abort();
 498 }
 499
 500 /* Disable signals, take mutex, remove from registry */
 501 static
 502 void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 503 {
 504         sigset_t newmask, oldmask;
 505         int ret;
 506
 507         ret = sigfillset(&newmask);
 508         if (ret)
 509                 abort();
 510         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 511         if (ret)
 512                 abort();
 513
 514         mutex_lock(&rcu_gp_lock);
 515         remove_thread(rcu_reader_reg);
 516         mutex_unlock(&rcu_gp_lock);
 517         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 518         if (ret)
 519                 abort();
 520         rcu_bp_exit();
 521 }
 522
 523 /*
 524  * Remove thread from the registry when it exits, and flag it as
 525  * destroyed so garbage collection can take care of it.
 526  */
 527 static
 528 void urcu_bp_thread_exit_notifier(void *rcu_key)
 529 {
 530         rcu_bp_unregister(rcu_key);
 531 }
 532
 533 static
 534 void rcu_bp_init(void)
 535 {
 536         mutex_lock(&init_lock);
 537         if (!rcu_bp_refcount++) {
 538                 int ret;
 539
 540                 ret = pthread_key_create(&urcu_bp_key,
 541                                 urcu_bp_thread_exit_notifier);
 542                 if (ret)
 543                         abort();
 544                 initialized = 1;
 545         }
 546         mutex_unlock(&init_lock);
 547 }
 548
 549 static
 550 void rcu_bp_exit(void)
 551 {
 552         mutex_lock(&init_lock);
 553         if (!--rcu_bp_refcount) {
 554                 struct registry_chunk *chunk, *tmp;
 555                 int ret;
 556
 557                 cds_list_for_each_entry_safe(chunk, tmp,
 558                                 &registry_arena.chunk_list, node) {
 559                         munmap(chunk, chunk->data_len
 560                                         + sizeof(struct registry_chunk));
 561                 }
 562                 ret = pthread_key_delete(urcu_bp_key);
 563                 if (ret)
 564                         abort();
 565         }
 566         mutex_unlock(&init_lock);
 567 }
 568
 569 /*
 570  * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
 571  * a concurrent thread executing with this same lock held. This ensures that the
 572  * registry is in a coherent state in the child.
 573  */
 574 void rcu_bp_before_fork(void)
 575 {
 576         sigset_t newmask, oldmask;
 577         int ret;
 578
 579         ret = sigfillset(&newmask);
 580         assert(!ret);
 581         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 582         assert(!ret);
 583         mutex_lock(&rcu_gp_lock);
 584         saved_fork_signal_mask = oldmask;
 585 }
 586
 587 void rcu_bp_after_fork_parent(void)
 588 {
 589         sigset_t oldmask;
 590         int ret;
 591
 592         oldmask = saved_fork_signal_mask;
 593         mutex_unlock(&rcu_gp_lock);
 594         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 595         assert(!ret);
 596 }
 597
 598 /*
 599  * Prune all entries from registry except our own thread. Fits the Linux
 600  * fork behavior. Called with rcu_gp_lock held.
 601  */
 602 static
 603 void urcu_bp_prune_registry(void)
 604 {
 605         struct registry_chunk *chunk;
 606         struct rcu_reader *rcu_reader_reg;
 607
 608         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 609                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 610                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 611                                 rcu_reader_reg++) {
 612                         if (!rcu_reader_reg->alloc)
 613                                 continue;
 614                         if (rcu_reader_reg->tid == pthread_self())
 615                                 continue;
 616                         cleanup_thread(chunk, rcu_reader_reg);
 617                 }
 618         }
 619 }
 620
 621 void rcu_bp_after_fork_child(void)
 622 {
 623         sigset_t oldmask;
 624         int ret;
 625
 626         urcu_bp_prune_registry();
 627         oldmask = saved_fork_signal_mask;
 628         mutex_unlock(&rcu_gp_lock);
 629         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 630         assert(!ret);
 631 }
 632
 633 void *rcu_dereference_sym_bp(void *p)
 634 {
 635         return _rcu_dereference(p);
 636 }
 637
 638 void *rcu_set_pointer_sym_bp(void **p, void *v)
 639 {
 640         cmm_wmb();
 641         uatomic_set(p, v);
 642         return v;
 643 }
 644
 645 void *rcu_xchg_pointer_sym_bp(void **p, void *v)
 646 {
 647         cmm_wmb();
 648         return uatomic_xchg(p, v);
 649 }
 650
 651 void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
 652 {
 653         cmm_wmb();
 654         return uatomic_cmpxchg(p, old, _new);
 655 }
 656
 657 DEFINE_RCU_FLAVOR(rcu_flavor);
 658
 659 #include "urcu-call-rcu-impl.h"
 660 #include "urcu-defer-impl.h"